Abstract
Background: Flat foot deformity is a prevalent and challenging condition often leading to various clinical complications. Accurate identification of abnormal foot types is essential for appropriate interventions. Method: A dataset consisting of 1573 plantar pressure images from 125 individuals was collected. The performance of the You Only Look Once v5 (YOLO-v5) model, improved YOLO-v5 model, and multi-label classification model was evaluated for foot type identification using the collected images. A new dataset was also collected to verify and compare the models. Results: The multi-label classification algorithm based on ResNet-50 outperformed other algorithms. The improved YOLO-v5 model with Squeeze-and-Excitation (SE), the improved YOLO-v5 model with Convolutional Block Attention Module (CBAM), and the multilabel classification model based on ResNet-50 achieved an accuracy of 0.652, 0.717, and 0.826, respectively, which is significantly higher than those obtained using the ordinary plantar-pressure system and the standard YOLO-v5 model. Conclusion: These results indicate that the proposed DL-based multilabel classification model based on ResNet-50 is superior in flat foot type detection and can be used to evaluate the clinical rehabilitation status of patients with abnormal foot types and various foot pathologies when more data on patients with various diseases are available for training.